A method and system for diagnosing faults of rotating machinery under variable working conditions

Abstract

The present invention discloses a method and system for diagnosing faults of rotating machinery under variable working conditions, which relates to the technical field of fault diagnosis, and includes the following steps: obtaining a labeled feature set of source domain data of rotating machinery and an unlabeled feature set of target domain data; performing cluster analysis on the labeled feature set to obtain the adjusted Rand index of each deep feature; calculating the multi-core maximum mean difference between each deep feature sample of the labeled feature set and the unlabeled feature set under normal working conditions; constructing a transferability index of each deep feature, and using the improved joint distribution adaptation IJDA for transfer learning, using the training feature set after transfer learning to train a recognition classifier, using the recognition classifier to identify and classify fault patterns, and outputting cross-domain fault diagnosis results. The present invention quantitatively evaluates the category discrimination of deep features and the distribution differences of features between different domains, which helps to select deep features that are more conducive to migration fault diagnosis, and has a high diagnostic accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of rotating machinery fault diagnosis under variable operating conditions, and in particular to a rotating machinery fault diagnosis method and system under variable operating conditions. Background Art

[0002] With the rapid development of a series of artificial intelligence methods such as machine learning, the fault diagnosis framework based on artificial intelligence has gradually become a research hotspot in data-driven rotating machinery fault diagnosis methods. At present, deep learning methods have attracted the attention and research of many researchers due to their powerful hidden feature mining capabilities, and have achieved many research results. However, due to the complex working conditions of rotating machinery in actual industrial scenarios, the fault diagnosis model based on deep learning faces two technical problems:

[0003] (1) There is a lack of sufficient labeled fault samples. In actual industrial scenarios, rotating machinery is subject to variable and complex working conditions, and there is a lack of sample data under different fault states. The cost of obtaining sufficient labeled samples is very high.

[0004] (2) Under different working conditions, there are distribution differences among samples of the same fault category, and most deep learning-based models are trained under the condition that the training samples and the test samples have the same distribution. As a result, the fault diagnosis model has poor fault identification effect on samples under different working conditions.

[0005] As a promising research direction to solve the above problems, transfer learning has gradually attracted the attention and research of researchers in recent years. Transfer learning methods can identify and classify fault samples from different fields (target domain: unlabeled fault samples under other working conditions) by mining learning knowledge and training fault diagnosis models from existing fields (source domain: labeled fault samples under existing working conditions). Although many researchers have used transfer learning methods to diagnose rotating machinery faults under different working conditions, these studies focus more on model migration or reducing inter-domain distribution differences, and lack comprehensive consideration of the domain invariance and discrimination performance of the extracted features, which have an important impact on improving the effectiveness and generalization ability of fault diagnosis models under different working conditions.

[0006] The migration fault diagnosis method based on deep learning in the prior art mainly includes four steps:

[0007] (1) Raw signal processing based on time-frequency analysis methods, which is to extract fault features from the raw vibration signals of mechanical equipment for subsequent fault diagnosis model training. Time-frequency analysis methods are often used to decompose vibration signals and extract time domain and source domain features to construct the original feature set. Commonly used time-frequency analysis methods include empirical mode decomposition, wavelet packet transform, short-time Fourier transform, etc.

[0008] (2) Pre-training of fault diagnosis models based on deep learning, which is to further mine hidden features from the original feature set. Deep learning methods are used to mine deep features from the original feature set, build a fully connected layer, and combine the Softmax classifier to obtain a fault diagnosis model. Commonly used deep learning methods include deep autoencoders, convolutional neural networks, deep belief networks, deep residual networks, etc.

[0009] (3) Migrate the fault diagnosis model and fine-tune it. Based on the fault diagnosis model constructed in the second step, the source domain samples are used to train the fault diagnosis model. Then, the obtained network parameters are migrated to the target domain fault diagnosis model, and the target domain fault diagnosis model is fine-tuned using the data in the normal state of the target domain to obtain the migrated fault diagnosis model.

[0010] (4) Fault pattern recognition and classification: Based on the migration fault diagnosis model obtained in the third step, the target domain samples are used as data to obtain the fault diagnosis results.

[0011] There are two main problems with the technical solutions in the prior art:

[0012] One-sided reliance on the hidden feature mining capabilities of deep learning methods without further analysis of the extracted deep features. Not all deep features extracted by deep learning methods are conducive to migration fault diagnosis. There may be interference and redundant features, which affect the accuracy of migration fault diagnosis.

[0013] The migration fault diagnosis model obtained by deep learning model parameter migration and fine-tuning does not fully consider the distribution differences between the source domain and the target domain when classifying fault samples, resulting in insufficient generalization ability of the fault diagnosis model. Summary of the invention

[0014] The purpose of the present invention is to provide a method and system for fault diagnosis of rotating machinery under variable working conditions, by further analyzing the deep features extracted by the deep learning method, selecting the deep features that are more conducive to migration fault diagnosis. In addition, it is proposed to use improved joint distribution adaptation to fully consider the distribution differences between data in different domains in the fault diagnosis process, so as to improve the generalization ability of the fault diagnosis model.

[0015] The present invention provides a method for diagnosing faults of rotating machinery under variable operating conditions, comprising the following steps:

[0016] Obtain source domain data of vibration signals of rotating machinery under known working conditions and target domain data of vibration signals under working conditions to be diagnosed, perform deep feature extraction on each of them, and obtain labeled feature sets of source domain data and unlabeled feature sets of target domain data;

[0017] Perform cluster analysis on the labeled feature set to obtain the adjusted Rand index (ARI) of each deep feature;

[0018] Calculate the multi-core maximum mean difference MK-MMD between each deep feature sample of the labeled feature set and the unlabeled feature set under normal working conditions;

[0019] Based on the adjusted Rand index ARI and multi-core maximum mean difference MK-MMD of each deep feature, the transferability index of each deep feature is constructed;

[0020] According to the transferability index of each deep feature, an improved joint distribution adaptation IJDA based on joint distribution adaptation JDA and introducing a balance factor in the optimization objective is used to perform transfer learning on the labeled feature set and the unlabeled feature set respectively to obtain the training feature set after transfer learning;

[0021] Construct a recognition classifier, use the training feature set after transfer learning to train the recognition classifier, use the trained recognition classifier to perform fault pattern recognition and classification on the vibration signal of the rotating machinery under the diagnosis condition, and output the cross-domain fault diagnosis results.

[0022] Furthermore, a deep autoencoder network based on different activation functions is used to perform deep feature extraction on the source domain data and the target domain data respectively, so as to obtain a labeled feature set of the source domain data and an unlabeled feature set of the target domain data, and construct a deep feature pool.

[0023] Furthermore, the K-means algorithm is used to perform cluster analysis on the labeled feature set to obtain the adjusted Rand index ARI of each deep feature.

[0024] Furthermore, the clustering analysis of the labeled feature set using the K-means algorithm includes:

[0025] When the labeled feature set of the source domain data is There are M features in total, among which the mth feature for:

[0026]

[0027] in is the nth feature sample of the kth fault category, k∈[1,K], n∈[1,N];

[0028] The K-means algorithm is used to obtain the adjusted Rand index ARI value of each feature sample. The adjusted Rand index ARI values ​​of M features construct a sequence as follows:

[0029] ARI={ari(1), ari(2),..., ari(m),..., ari(M)} (2)

[0030] Where ari(m) is the adjusted Rand index of the mth feature.

[0031] Furthermore, constructing the transferability index of each deep feature according to the adjusted Rand index ARI and the multi-core maximum mean difference MK-MMD of each deep feature includes the following steps:

[0032] The transferability index of each feature is calculated separately, and the calculation formula is:

[0033]

[0034] Then the transferability index values ​​of M features construct a sequence as follows:

[0035] RAM={ram(1), ram(2),..., ram(m),..., ram(M)} (4)

[0036] Where mkmmd(m) is the multi-core maximum mean difference value of the mth feature;

[0037] ram(m) is the transferability index value of the mth feature.

[0038] Furthermore, the transferability index based on each deep feature adopts an improved joint distribution adaptation IJDA based on joint distribution adaptation JDA and introducing a balance factor in the optimization objective to perform transfer learning on the labeled feature set and the unlabeled feature set respectively, including:

[0039] When there is a labeled feature set D S ={(x1,y1),...,(x ns ,y ns )}, the unlabeled feature set is Based on the labeled feature set D S With the unlabeled feature set D T , and obtain a mapping transformation A;

[0040] Among them, n s is the number of source domain samples; x ns is the training sample; ns is the category label corresponding to the source domain sample;

[0041] n T is the number of samples in the target domain; For nth s +n T Test samples;

[0042] The metric expression of the marginal probability distribution of the labeled feature set and the unlabeled feature set after A mapping transformation is as follows:

[0043]

[0044] Among them, M0 is the maximum mean difference matrix, and its expression is as follows:

[0045]

[0046] Among them, (M0) ij is the maximum mean difference matrix between the i-th source domain sample and the j-th target domain sample;

[0047] i and j represent the i-th source domain sample and the j-th target domain sample respectively;

[0048] x i is the i-th source domain sample; x j is the jth target domain sample;

[0049] The maximum mean difference distance of the conditional probability distribution of the marginal probability distribution of the labeled feature set and the unlabeled feature set after A mapping transformation is:

[0050]

[0051] Among them, C is the number of sample categories;

[0052] is the c-th sample in the target domain;

[0053] is the c-th sample in the source domain;

[0054] is the number of samples of the cth class in the target domain; H is the reproducing kernel Hilbert space;

[0055] A T x S is the source domain feature sample after being transformed by matrix A;

[0056] A T x T is the target domain feature sample after being transformed by matrix A;

[0057] M c To consider the maximum mean difference matrix of sample categories, its expression is as follows:

[0058]

[0059] The overall optimization goal of the improved joint distribution adaptation IJDA is:

[0060]

[0061] st A T XHXT A=I (9)

[0062] in, is the regularization term;

[0063] A T XHX T A=I is a constraint condition;

[0064] α is the balance factor, α∈[0,1].

[0065] Furthermore, the training method for training the recognition classifier using the training feature set after transfer learning includes support vector machine, K nearest neighbor, and random forest method.

[0066] The present invention provides a rotating machinery variable operating condition fault diagnosis system, comprising:

[0067] The data set acquisition module is used to acquire the source domain data of the vibration signal of the rotating machinery under known working conditions and the target domain data of the vibration signal under the working conditions to be diagnosed, and perform deep feature extraction respectively to obtain the labeled feature set of the source domain data and the unlabeled feature set of the target domain data;

[0068] The deep feature extraction module is used to perform cluster analysis on the labeled feature set to obtain the adjusted Rand index (ARI) of each deep feature;

[0069] The mean difference value acquisition module calculates the multi-core maximum mean difference MK-MMD between each deep feature sample of the labeled feature set and the unlabeled feature set under normal working conditions;

[0070] The transferability index acquisition module is used to construct the transferability index of each deep feature according to the adjusted Rand index ARI and the multi-core maximum mean difference MK-MMD of each deep feature;

[0071] The training data set acquisition module is used to perform transfer learning on the labeled feature set and the unlabeled feature set respectively according to the transferability index of each deep feature, using the improved joint distribution adaptation IJDA based on the joint distribution adaptation JDA and introducing the balance factor in the optimization objective, to obtain the training feature set after transfer learning;

[0072] The classifier training module is used to build a recognition classifier and train the recognition classifier using the training feature set after transfer learning;

[0073] The diagnosis result acquisition module uses the trained recognition classifier to perform fault pattern recognition and classification on the vibration signal of the rotating machinery under the diagnosis condition, and outputs the cross-domain fault diagnosis result.

[0074] Compared with the prior art, the present invention has the following beneficial effects:

[0075] (1) The present invention can quantitatively evaluate the category discrimination of deep features and the distribution differences of features between different domains, which helps to select deep features that are more conducive to migration fault diagnosis and remove interference and redundant deep features. When an appropriate number of deep features are selected, ideal fault diagnosis results can be obtained with high diagnostic accuracy.

[0076] (2) The deep feature selection method for cross-domain diagnosis proposed in the present invention is combined with the feature transfer learning method. The selected features with high category discrimination and domain invariance can significantly improve the performance of the model's cross-domain fault diagnosis and improve the accuracy of cross-domain diagnosis.

[0077] (3) Compared with the classical feature-based transfer learning methods (transfer component analysis, joint distribution adaptation), the improved joint distribution adaptation IJDA feature transfer learning method proposed in the present invention has a better ability to reduce the inter-domain distribution differences and promote the improvement of the generalization ability of the fault diagnosis model. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0079] Figure 1 It is a flowchart of a method for diagnosing faults of rotating machinery under variable working conditions proposed by the present invention;

[0080] Figure 2 It is a detailed flow chart of a method for diagnosing faults of rotating machinery under variable working conditions proposed by the present invention;

[0081] Figure 3 It is a physical picture of the SQI-MFS mechanical comprehensive fault simulation test bench in an embodiment of a rotating machinery variable operating condition fault diagnosis method proposed by the present invention. DETAILED DESCRIPTION

[0082] The technical solutions in the embodiments of the present invention are clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention, but it should be understood that the protection scope of the present invention is not limited by the specific implementation methods.

[0083] Example

[0084] like Figure 1-2 As shown, a method for diagnosing faults of rotating machinery under variable working conditions comprises the following steps:

[0085] Step 1: Obtain the source domain data of the vibration signal of the rotating machinery under known working conditions and the target domain data of the vibration signal under the working conditions to be diagnosed, perform deep feature extraction on each of them, and obtain the labeled feature set of the source domain data and the unlabeled feature set of the target domain data.

[0086] By using a deep autoencoder network based on different activation functions to extract deep features of the source domain data and the target domain data respectively, we can obtain a labeled feature set of the source domain data and an unlabeled feature set of the target domain data, and construct a deep feature pool.

[0087] Step 2: Perform cluster analysis on the labeled feature set to obtain the adjusted Rand index (ARI) of each deep feature;

[0088] In order to select deep features that are more beneficial to transfer learning from the deep feature pool, a deep feature selection method TFSAM for transfer learning is proposed. The deep features are quantitatively evaluated from two perspectives: the discriminative performance and domain invariance of the features. TFSAM for cross-domain diagnosis includes the following steps:

[0089] The K-means algorithm is used to perform cluster analysis on the labeled feature set to obtain the adjusted Rand index (ARI) of each deep feature, including:

[0090] When the labeled feature set of the source domain data is There are M features in total, among which the mth feature for:

[0091]

[0092] in is the nth feature sample of the kth fault category, k∈[1,K], n∈[1,N];

[0093] The K-means algorithm is used to obtain the adjusted Rand index ARI value of each feature sample. The adjusted Rand index ARI values ​​of M features construct a sequence as follows:

[0094] ARI={ari(1), ari(2),..., ari(m),..., ari(M)} (2)

[0095] Where ari(m) is the adjusted Rand index of the mth feature.

[0096] The larger the adjusted Rand Index (ARI) value of a feature, the better the category discrimination of the feature.

[0097] Step 3: Calculate the multi-core maximum mean difference MK-MMD between each deep feature sample of the labeled feature set and the unlabeled feature set under normal working conditions.

[0098] The feature cross-domain invariance evaluation based on the multi-core maximum mean difference MK-MMD is performed on the feature samples extracted from the labeled feature set and the unlabeled feature set under the normal state of the device. The multi-core maximum mean difference MK-MMD value of the features between the two domains is calculated to quantify the cross-domain invariance of the features.

[0099] Then the multi-core maximum mean difference MK-MMD value of M features constructs a sequence as:

[0100] MKMMD={mkmmd(1), mkmmd(2), mkmmd(m),..., mkmmd(M)} (3)

[0102] Where mkmmd(m) is the multi-core maximum mean difference value of the mth feature.

[0103] Step 4: Based on the adjusted Rand index ARI and the multi-core maximum mean difference MK-MMD of each deep feature, the transferability index of each deep feature is constructed, including the following steps:

[0104] The transferability index of each feature is calculated separately, and the calculation formula is:

[0105]

[0106] Then the transferability index values ​​of M features construct a sequence as follows:

[0107] RAM={ram(1), ram(2),..., ram(m),..., ram(M)} (5)

[0108] Where ram(m) is the transferability index value of the mth feature.

[0109] The larger the transferability index value of a feature is, the better the transferability of the feature is.

[0110] The larger the feature transferability index value RAM, the better the feature's category discrimination, that is, the higher the discrimination performance. At the same time, the smaller the distribution difference under different working conditions, that is, the better the cross-domain invariance. Therefore, the feature with a large feature transferability index value RAM has a higher value for subsequent transfer learning, which is more conducive to improving the transfer learning effect, and the better the cross-domain recognition and classification performance of the trained fault pattern recognition classifier. Therefore, the RAM values ​​of the M features are sorted by size, and the features with large RAM values ​​are selected for the next step of feature transfer learning.

[0111] Step 5: According to the transferability index of each deep feature, the improved joint distribution adaptation IJDA based on joint distribution adaptation JDA and introducing a balance factor in the optimization objective is used to perform transfer learning on the labeled feature set and the unlabeled feature set respectively to obtain the training feature set after transfer learning.

[0112] Compared with joint distribution adaptation JDA, improved joint distribution adaptation IJDA has two improvements:

[0113] ① Balance the marginal probability and conditional probability distribution, introduce a balance factor into the optimization objective, and achieve adaptive adjustment of the two distributions;

[0114] ② In order to further improve the effect of transfer learning and adaptive efficiency, the multi-core maximum distribution difference is used as a measurement method for data distribution difference.

[0115] By performing transfer learning on the feature sets of the source domain and the target domain, the distribution differences between the domains are reduced, and the training set and test set after transfer learning are obtained for subsequent training and testing of the fault mode recognition classifier.

[0116] Based on the transferability index of each deep feature, an improved joint distribution adaptation IJDA based on joint distribution adaptation JDA and introducing a balance factor in the optimization objective is used to perform transfer learning on the labeled feature set and the unlabeled feature set, respectively, including:

[0117] When there is a labeled feature set D S ={(x1,y1),...,(x ns ,y ns )}, the unlabeled feature set is Based on the labeled feature set D S With the unlabeled feature set D T , and obtain a mapping transformation A;

[0118] Among them, n s is the number of source domain samples; x ns is the training sample; ns is the category label corresponding to the source domain sample;

[0119] n T is the number of samples in the target domain; For nth s +n T Test samples;

[0120] Assume that the marginal probability distribution and conditional probability distribution of the two domain data sets are not equal, that is, P S (x S )≠P T (x T ) and Q S (yS |x S )≠Q T (y T |x T ).

[0121] The goal of the JDA algorithm is to S With D T The sample learns a mapping transformation A.

[0122] The metric expression of the marginal probability distribution of the labeled feature set and the unlabeled feature set after A mapping transformation is as follows:

[0123]

[0124] Among them, M0 is the maximum mean difference matrix, and its expression is as follows:

[0125]

[0126] Among them, (M0) ij is the maximum mean difference matrix between the i-th source domain sample and the j-th target domain sample;

[0127] i and j represent the i-th source domain sample and the j-th target domain sample respectively;

[0128] x i is the i-th source domain sample; x j is the jth target domain sample;

[0129] The maximum mean difference distance of the conditional probability distribution of the marginal probability distribution of the labeled feature set and the unlabeled feature set after A mapping transformation is:

[0130]

[0131] Among them, C is the number of sample categories;

[0132] is the c-th sample in the target domain;

[0133] is the c-th sample in the source domain;

[0134] is the number of samples of the cth class in the target domain; H is the reproducing kernel Hilbert space;

[0135] A T x S is the source domain feature sample after being transformed by matrix A;

[0136] A T x Tis the target domain feature sample after being transformed by matrix A;

[0137] M c To consider the maximum mean difference matrix of sample categories, its expression is as follows:

[0138]

[0139] The overall optimization goal of the improved joint distribution adaptation IJDA is:

[0140]

[0141] st A T XHX T A=I (10)

[0142] in, is the regularization term;

[0143] A T XHX T A=I is a constraint condition;

[0144] α is the balance factor, α∈[0,1], and its value is determined based on manual experience.

[0145] Step 6: Construct a recognition classifier, use the training feature set after transfer learning to train the recognition classifier, use the trained recognition classifier to perform fault pattern recognition and classification on the vibration signal of the rotating machinery under the diagnosis condition, and output the cross-domain fault diagnosis result.

[0146] The training methods for training recognition classifiers using the training feature set after transfer learning include support vector machine, K nearest neighbor, and random forest method.

[0147] Example 2

[0148] The present invention provides a rotating machinery variable operating condition fault diagnosis system, comprising:

[0149] The data set acquisition module is used to acquire the source domain data of the vibration signal of the rotating machinery under known working conditions and the target domain data of the vibration signal under the working conditions to be diagnosed, and perform deep feature extraction respectively to obtain the labeled feature set of the source domain data and the unlabeled feature set of the target domain data;

[0150] The deep feature extraction module is used to perform cluster analysis on the labeled feature set to obtain the adjusted Rand index (ARI) of each deep feature;

[0151] The mean difference value acquisition module calculates the multi-core maximum mean difference MK-MMD between each deep feature sample of the labeled feature set and the unlabeled feature set under normal working conditions;

[0152] The transferability index acquisition module is used to construct the transferability index of each deep feature according to the adjusted Rand index ARI and the multi-core maximum mean difference MK-MMD of each deep feature;

[0153] The training data set acquisition module is used to perform transfer learning on the labeled feature set and the unlabeled feature set respectively according to the transferability index of each deep feature, using the improved joint distribution adaptation IJDA based on the joint distribution adaptation JDA and introducing the balance factor in the optimization objective, to obtain the training feature set after transfer learning;

[0154] The classifier training module is used to build a recognition classifier and train the recognition classifier using the training feature set after transfer learning;

[0155] The diagnosis result acquisition module uses the trained recognition classifier to perform fault pattern recognition and classification on the vibration signal of the rotating machinery under the diagnosis condition, and outputs the cross-domain fault diagnosis result.

[0156] The present invention will be further described below in conjunction with specific implementation data.

[0157] like Figure 3 The present invention utilizes the SQI-MFS mechanical comprehensive fault simulation test bench to collect bearing and motor fault vibration signals at speeds of 1200r / min and 1800r / min, respectively.

[0158] Among them, the bearing fault diagnosis experimental data set is shown in Table 1, and the motor fault diagnosis experimental data set is shown in Table 2.

[0159] Table 1 Bearing fault diagnosis experimental data set

[0160]

[0161] Table 2 Motor fault diagnosis experimental data set

[0162]

[0163] Based on the bearing fault diagnosis experimental data set shown in Table 1 and the motor fault diagnosis experimental data set shown in Table 2, using the rotating machinery variable operating condition fault diagnosis method of the present invention, the accuracy of bearing fault diagnosis can reach 92.92%, and the accuracy of motor fault diagnosis can reach 95.42%.

[0164] The present invention can quantitatively evaluate the category discrimination of deep features and the distribution differences of features between different domains, which helps to select deep features that are more conducive to migration fault diagnosis and remove interfering and redundant deep features. When an appropriate number of deep features are selected, ideal fault diagnosis results can be obtained.

[0165] Finally, it should be noted that what is disclosed above is only a specific embodiment of the present invention, but the embodiment of the present invention is not limited thereto, and any changes that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for diagnosing faults of rotating machinery under variable operating conditions, characterized in that: The following steps are involved: Obtain source domain data of vibration signals of rotating machinery under known working conditions and target domain data of vibration signals under working conditions to be diagnosed, perform deep feature extraction on each of them, and obtain labeled feature sets of source domain data and unlabeled feature sets of target domain data; Perform cluster analysis on the labeled feature set to obtain the adjusted Rand index (ARI) of each deep feature; Calculate the multi-core maximum mean difference MK-MMD between each deep feature sample of the labeled feature set and the unlabeled feature set under normal working conditions; Based on the adjusted Rand index ARI and multi-core maximum mean difference MK-MMD of each deep feature, the transferability index of each deep feature is constructed; According to the transferability index of each deep feature, an improved joint distribution adaptation IJDA based on joint distribution adaptation JDA and introducing a balance factor in the optimization objective is used to perform transfer learning on the labeled feature set and the unlabeled feature set respectively to obtain the training feature set after transfer learning; Construct a recognition classifier, use the training feature set after transfer learning to train the recognition classifier, use the trained recognition classifier to perform fault pattern recognition and classification on the vibration signal of the rotating machinery under the working condition to be diagnosed, and output the cross-domain fault diagnosis results; Among them, according to the transferability index of each deep feature, the improved joint distribution adaptation IJDA based on joint distribution adaptation JDA and introducing a balance factor in the optimization target is used to perform transfer learning on the labeled feature set and the unlabeled feature set respectively, including: When there is a labeled feature set D S ={(x1,y1),...,(x ns ,y ns )}, the unlabeled feature set is Based on the labeled feature set D S With the unlabeled feature set D T , and obtain a mapping transformation A; Among them, n s is the number of source domain samples; x ns is the training sample; ns is the category label corresponding to the source domain sample; n T is the number of samples in the target domain; For nth s +n T Test samples; The metric expression of the marginal probability distribution of the labeled feature set and the unlabeled feature set after A mapping transformation is as follows: Where M0 is the maximum mean difference matrix, which is expressed as follows: Among them, (M0) ij is the maximum mean difference matrix between the i-th source domain sample and the j-th target domain sample; i and j represent the i-th source domain sample and the j-th target domain sample respectively; x i is the i-th source domain sample; x j is the jth target domain sample; The maximum mean difference distance of the conditional probability distribution of the marginal probability distribution of the labeled feature set and the unlabeled feature set after A mapping transformation is: Among them, C is the number of sample categories; is the c-th sample in the target domain; is the c-th sample in the source domain; is the number of samples of the cth class in the target domain; H is the reproducing kernel Hilbert space; A T x S is the source domain feature sample after being transformed by matrix A; A T x T is the target domain feature sample after being transformed by matrix A; M c To consider the maximum mean difference matrix of sample categories, its expression is as follows: The overall optimization goal of the improved joint distribution adaptation IJDA is: s.t.A T XHX T A=I in, is the regularization term; A T XHX T A=I is a constraint condition; α is the balance factor, α∈[0,1].

2. A method for diagnosing faults of rotating machinery under variable operating conditions according to claim 1, characterized in that: A deep autoencoder network based on different activation functions is used to perform deep feature extraction on the source domain data and the target domain data respectively, to obtain a labeled feature set of the source domain data and an unlabeled feature set of the target domain data, and to construct a deep feature pool.

3. A method for diagnosing faults of rotating machinery under variable operating conditions according to claim 2, characterized in that: The K-means algorithm is used to perform cluster analysis on the labeled feature set to obtain the adjusted Rand index ARI of each deep feature.

4. A method for diagnosing faults of rotating machinery under variable operating conditions according to claim 3, characterized in that: The clustering analysis of the labeled feature set using the K-means algorithm includes: When the labeled feature set of the source domain data is There are M features in total, among which the mth feature for: in is the nth feature sample of the kth fault category, k∈[1,K], n∈[1,N]; The K-means algorithm is used to obtain the adjusted Rand index ARI value of each feature sample. The adjusted Rand index ARI values ​​of M features construct a sequence as follows: ARI={ari(1), ari(2),..., ari(m),..., ari(M)} (2) Where ari(m) is the adjusted Rand index of the mth feature.

5. A method for diagnosing faults of rotating machinery under variable operating conditions according to claim 4, characterized in that: The method of constructing the transferability index of each deep feature according to the adjusted Rand index ARI and the multi-core maximum mean difference MK-MMD of each deep feature includes the following steps: The transferability index of each feature is calculated separately, and the calculation formula is: Then the transferability index values ​​of M features construct a sequence as follows: RAM={ram(1), ram(2),..., ram(m),..., ram(M)} (4) Where mkmmd(m) is the multi-core maximum mean difference value of the mth feature; ram(m) is the transferability index value of the mth feature.

6. A method for diagnosing faults of rotating machinery under variable operating conditions according to claim 1, characterized in that: The training method for training the recognition classifier using the training feature set after transfer learning includes support vector machine, K nearest neighbor, and random forest method.

7. A rotating machinery variable operating condition fault diagnosis system, characterized in that: include: The data set acquisition module is used to acquire the source domain data of the vibration signal of the rotating machinery under known working conditions and the target domain data of the vibration signal under the working conditions to be diagnosed, and perform deep feature extraction respectively to obtain the labeled feature set of the source domain data and the unlabeled feature set of the target domain data; The deep feature extraction module is used to perform cluster analysis on the labeled feature set to obtain the adjusted Rand index (ARI) of each deep feature; The mean difference value acquisition module calculates the multi-core maximum mean difference MK-MMD between each deep feature sample of the labeled feature set and the unlabeled feature set under normal working conditions; The transferability index acquisition module is used to construct the transferability index of each deep feature according to the adjusted Rand index ARI and the multi-core maximum mean difference MK-MMD of each deep feature; The training data set acquisition module is used to perform transfer learning on the labeled feature set and the unlabeled feature set respectively according to the transferability index of each deep feature, using the improved joint distribution adaptation IJDA based on the joint distribution adaptation JDA and introducing the balance factor in the optimization objective, to obtain the training feature set after transfer learning; The training data set acquisition module includes: When there is a labeled feature set D s ={(x1,y1),...,(x ns ,y ns )}, the unlabeled feature set is Based on the labeled feature set D S With the unlabeled feature set D T , and obtain a mapping transformation A; Among them, n s is the number of source domain samples; x ns is the training sample; ns is the category label corresponding to the source domain sample; n T is the number of samples in the target domain; For nth s +n T Test samples; The metric expression of the marginal probability distribution of the labeled feature set and the unlabeled feature set after A mapping transformation is as follows: Where M0 is the maximum mean difference matrix, which is expressed as follows: Among them, (M0) ij is the maximum mean difference matrix between the i-th source domain sample and the j-th target domain sample; i and j represent the i-th source domain sample and the j-th target domain sample respectively; x i is the i-th source domain sample; x j is the jth target domain sample; The maximum mean difference distance of the conditional probability distribution of the marginal probability distribution of the labeled feature set and the unlabeled feature set after A mapping transformation is: Among them, C is the number of sample categories; is the c-th sample in the target domain; is the c-th sample in the source domain; is the number of samples of the cth class in the target domain; H is the reproducing kernel Hilbert space; A T x S is the source domain feature sample after being transformed by matrix A; A T x T is the target domain feature sample after being transformed by matrix A; M c To consider the maximum mean difference matrix of sample categories, its expression is as follows: The overall optimization goal of the improved joint distribution adaptation IJDA is: s.t.A T XHX T A=I in, is the regularization term; A T XHX T A=I is a constraint condition; α is the balance factor, α∈[0,1]; The classifier training module is used to build a recognition classifier and train the recognition classifier using the training feature set after transfer learning; The diagnosis result acquisition module uses the trained recognition classifier to perform fault pattern recognition and classification on the vibration signal of the rotating machinery under the diagnosis condition, and outputs the cross-domain fault diagnosis result.

Images